Strix Halo' iGPU is RDNA3.5 and I'm not sure I'd buy an iGPU that can't do DLSS-like quality upscaling. Only the RDNA4 architecture can do that hardware-based ML upscaling (tho, NVIDIA's transformer model provides even better overall quality).
One could also see it this way: The 256-bit Strix Halo APU is using 8000 MT/s memory, but now, 9600 MT/s memory is available and is being used (by APPLE's M5 APU and INTEL's Arc B390). If Strix Halo would use the faster 9600 MT/s memory and is not limited by the memory bandwidth, then it could be up to 20% faster (9600/8000). So, maybe not using 9600 MT/s memory is hurting Strix Halo's per/W.